Since 1986 - Covering the Fastest Computers in the World and the People Who Run Them

Language Flags
August 10, 2007

Ethernet Headed for 40/100GbE Standard

by Michael Feldman

As Yogi Berra once said, “When you come to a fork in the road, take it.” That’s exactly what the IEEE 802.3 Higher Speed Study Group (HSSG) decided to do last month when it was confronted with two different paths to the future of Ethernet. In a nutshell, the study group recommended that the Ethernet standard move forward on two parallel tracks — a 40 Gigabit Ethernet (40GbE) data rate for server and storage applications and a 100 Gigabit Ethernet (100GbE) data rate for long-haul networking and network aggregation.

Once those recommendations are formally approved by the IEEE 802.3 executive committee and the IEEE standards board, the HSSG will morph into the task force, which will then develop the new standard under a unified effort. Approval is anticipated around November or December, with the task force beginning its work in January 2008. The new 40GbE and 100 GbE specification should be completed by mid-2010, which is expected to be followed very quickly by the first round of 40GbE and 100GbE products.

If the task force is able to keep to that schedule, it will have to overcome a sluggish start by the study group. The HSSG was originally formed in July 2006 and up until July 2007 had been mired in debate over about how Ethernet technology should evolve. There was even some media speculation that discussions were at an impasse and the whole process might need to be rebooted.

The Ethernet rate step has been the central issue to the ongoing debate in HSSG for the discussions started in earnest six months ago. The result is essentially a compromise, recognizing that the communication needs of the computing sector are different than that of the networking sector. The 40GbE rate is intended for shorter range interconnects involved in server and storage communications; the 100GbE for long-haul networking and aggregation. The implication is that a single implementation can support either speed, but not both.

The underlying protocol for both speeds is, of course, Ethernet. So for those interested in datacenter consolidation around a single communication protocol, the two speeds offer a relatively painless way to get that done.

According to HSSG chair John D’Ambrosia, the discussions over the past six months have been contentious. Apparently, it took some time for vendors in the two application areas to come to terms with each other. In the end, though, many of the decisions were reached unanimously.

“Was I surprised that the consensus was this high?” asks D’Ambrosia. “Given all the work that was done — no. But looking at where we were in May, it certainly represents a turnaround. I have the highest respect for the study group and those individuals who participated in this process. I knew that group would come to a decision so we could move forward.”

The proposed objectives for 40GbE applications will support:

  1. At least one meter over a backplane,
  2. At least 10m over a copper cable assembly, and
  3. At least 100m over OM3 (optical multimode 3) fiber.

The proposed objectives for 100GbE applications will support:

  1. At least 10m over copper cable assembly,
  2. At at least 100m over OM3 fiber, and
  3. Up to 40km over single-mode optical fiber.

One of the factors that influenced the dual approach was the realization that computing and networking bandwidth demand are growing at different rates. Server-based bandwidth demand is doubling roughly every 24 months, while networking bandwidth demand is doubling approximately every 18 months. Because of the different trajectories, the Ethernet study group concluded that the server space won’t be ready for 100GbE for another several years. They believe that the 40GbE rate will match up with the next generation of multicore processors, memory technologies, and host bus interfaces. In the same timeframe, 100GbE networking will be required by the evolution of the Internet and the associated datacenter buildout.

But 40GbE may be behind the curve in the HPC sector, where performance is often limited by the relatively slow node-to-node interconnects. InfiniBand usage is growing in HPC systems and shows no signs of reversing. It remains to be seen how 40GbE will play against the more aggressive roadmap of the InfiniBand vendors, who are already laying the groundwork for products that will support 120 Gbps.

What Ethernet has on its side is ubiquity. But at these higher data rates, the speed of individual communication lanes and power consumption will present technical challenges for vendors. Fortunately, the implementation of 100GbE offers some options. One can vary the number of lanes depending on the speed rate for the individual lanes (e.g., 10*10GbE 4*25GbE, etc.). Cost will be the other big technical challenge at these higher speeds. As D’Ambrosia observes: “Ethernet likes to be cheap.”